Thinking Machines

AI Models · Open-Weights · Foundation Models

Launch

Thinking Machines ships Inkling-Small, matching its 975B model at 12B active

The smaller model beats its own larger sibling on coding and agentic benchmarks, undercutting the case for scale alone.

• Thinking Machines released Inkling-Small on July 30, 2026, an open-weights Mixture-of-Experts model with 276B total parameters and only 12B active per token, under an Apache 2.0 license.
• Thinking Machines Lab was founded by former OpenAI CTO Mira Murati and previously raised roughly $2 billion at a $12 billion valuation in a seed round led by Andreessen Horowitz.
• The 42-layer decoder routes each token to 6 of 256 specialized experts plus 2 always-on shared experts, and natively handles text, image and audio input with a context window up to 1M tokens.
• Inkling-Small scored 40 on Artificial Analysis's Intelligence Index versus 41 for the 975B-parameter Inkling, while beating it 80.2% to 77.6% on SWE-bench Verified.
• Inkling-Small arrived two weeks after the 975B-parameter flagship Inkling and was post-trained using on-policy distillation with Inkling itself as the teacher model.
• Shipping a smaller model that nearly matches its own flagship signals a shift from pure scale toward efficiency, and gives US labs a competitive open-weights entrant against a field long dominated by Chinese releases.